Mirror of the gdb-patches mailing list
 help / color / mirror / Atom feed
From: Yang Liu <liuyang22@iscas.ac.cn>
To: Palmer Dabbelt <palmer@dabbelt.com>, gdb-patches@sourceware.org
Cc: Andrew Burgess <aburgess@redhat.com>, simon.marchi@polymtl.ca
Subject: Re: [PATCH] gdb: RISC-V: Refine lr/sc sequence support
Date: Fri, 3 Nov 2023 01:54:38 +0800	[thread overview]
Message-ID: <bcd45a50-ec51-49fe-ae93-c19226f28d4a@iscas.ac.cn> (raw)
In-Reply-To: <mhng-e5f786bc-2a32-4288-958c-0743887d3d2f@palmer-ri-x1c9>


On 2023/11/2 23:58, Palmer Dabbelt wrote:
> On Thu, 02 Nov 2023 08:21:16 PDT (-0700), liuyang22@iscas.ac.cn wrote:
>>
>> On 2023/11/2 22:03, Yang Liu wrote:
>>> Per RISC-V spec, the lr/sc sequence can consist of up to 16 
>>> instructions, and we
>>> cannot insert breakpoints in the middle of this sequence. Before 
>>> this, we only
>>> detected a specific pattern (the most common one). This patch 
>>> improves this part
>>> and now supports more complex pattern detection.
>
> Thanks.  The general idea here seems pretty reasonable, but I haven't 
> looked that closely at the code.  I'm super behind on Linux stuff 
> right now, so I might be slower that usual...
>
No rush, thank you for your time!

>>> gdb/ChangeLog:
>>>
>>>          * gdb/riscv-tdep.c (class riscv_insn): Add more needed 
>>> opcode enums.
>>>          (riscv_insn::decode): Decode newly added opcodes.
>>> (riscv_insn_is_non_cti_and_allowed_in_atomic_sequence): New.
>>>          (riscv_insn_is_direct_branch): New.
>>>          (riscv_next_pc_atomic_sequence): Removed.
>>>          (riscv_deal_with_atomic_sequence): Rename from 
>>> riscv_next_pc_atomic_sequence.
>>>          (riscv_software_single_step): Adjust to use the renamed one.
>>>
>>> Signed-off-by: Yang Liu <liuyang22@iscas.ac.cn>
>>> ---
>>>   gdb/riscv-tdep.c | 266 
>>> ++++++++++++++++++++++++++++++++++++++++-------
>>>   1 file changed, 226 insertions(+), 40 deletions(-)
>>>
>>> diff --git a/gdb/riscv-tdep.c b/gdb/riscv-tdep.c
>>> index 3a2891c2c92..faff1aab420 100644
>>> --- a/gdb/riscv-tdep.c
>>> +++ b/gdb/riscv-tdep.c
>>> @@ -1578,8 +1578,34 @@ class riscv_insn
>>>         BLTU,
>>>         BGEU,
>>>         /* These are needed for stepping over atomic sequences.  */
>>> -      LR,
>>> -      SC,
>>> +      SLTI,
>>> +      SLTIU,
>>> +      XORI,
>>> +      ORI,
>>> +      ANDI,
>>> +      SLLI,
>>> +      SLLIW,
>>> +      SRLI,
>>> +      SRLIW,
>>> +      SRAI,
>>> +      SRAIW,
>>> +      SUB,
>>> +      SUBW,
>>> +      SLL,
>>> +      SLLW,
>>> +      SLT,
>>> +      SLTU,
>>> +      XOR,
>>> +      SRL,
>>> +      SRLW,
>>> +      SRA,
>>> +      SRAW,
>>> +      OR,
>>> +      AND,
>>> +      LR_W,
>>> +      LR_D,
>>> +      SC_W,
>>> +      SC_D,
>>>         /* This instruction is used to do a syscall.  */
>>>         ECALL,
>>>
>>> @@ -1768,6 +1794,13 @@ class riscv_insn
>>>       m_imm.s = EXTRACT_CBTYPE_IMM (ival);
>>>     }
>>>
>>> +  void decode_ca_type_insn (enum opcode opcode, ULONGEST ival)
>>> +  {
>>> +    m_opcode = opcode;
>>> +    m_rs1 = decode_register_index_short (ival, OP_SH_CRS1S);
>>> +    m_rs2 = decode_register_index_short (ival, OP_SH_CRS2S);
>>> +  }
>>> +
>>>     /* Fetch instruction from target memory at ADDR, return the 
>>> content of
>>>        the instruction, and update LEN with the instruction length.  */
>>>     static ULONGEST fetch_instruction (struct gdbarch *gdbarch,
>>> @@ -1882,14 +1915,62 @@ riscv_insn::decode (struct gdbarch *gdbarch, 
>>> CORE_ADDR pc)
>>>       decode_b_type_insn (BLTU, ival);
>>>         else if (is_bgeu_insn (ival))
>>>       decode_b_type_insn (BGEU, ival);
>>> +      else if (is_slti_insn(ival))
>>> +    decode_i_type_insn (SLTI, ival);
>>> +      else if (is_sltiu_insn(ival))
>>> +    decode_i_type_insn (SLTIU, ival);
>>> +      else if (is_xori_insn(ival))
>>> +    decode_i_type_insn (XORI, ival);
>>> +      else if (is_ori_insn(ival))
>>> +    decode_i_type_insn (ORI, ival);
>>> +      else if (is_andi_insn(ival))
>>> +    decode_i_type_insn (ANDI, ival);
>>> +      else if (is_slli_insn(ival))
>>> +    decode_i_type_insn (SLLI, ival);
>>> +      else if (is_slliw_insn(ival))
>>> +    decode_i_type_insn (SLLIW, ival);
>>> +      else if (is_srli_insn(ival))
>>> +    decode_i_type_insn (SRLI, ival);
>>> +      else if (is_srliw_insn(ival))
>>> +    decode_i_type_insn (SRLIW, ival);
>>> +      else if (is_srai_insn(ival))
>>> +    decode_i_type_insn (SRAI, ival);
>>> +      else if (is_sraiw_insn(ival))
>>> +    decode_i_type_insn (SRAIW, ival);
>>> +      else if (is_sub_insn(ival))
>>> +    decode_r_type_insn (SUB, ival);
>>> +      else if (is_subw_insn(ival))
>>> +    decode_r_type_insn (SUBW, ival);
>>> +      else if (is_sll_insn(ival))
>>> +    decode_r_type_insn (SLL, ival);
>>> +      else if (is_sllw_insn(ival))
>>> +    decode_r_type_insn (SLLW, ival);
>>> +      else if (is_slt_insn(ival))
>>> +    decode_r_type_insn (SLT, ival);
>>> +      else if (is_sltu_insn(ival))
>>> +    decode_r_type_insn (SLTU, ival);
>>> +      else if (is_xor_insn(ival))
>>> +    decode_r_type_insn (XOR, ival);
>>> +      else if (is_srl_insn(ival))
>>> +    decode_r_type_insn (SRL, ival);
>>> +      else if (is_srlw_insn(ival))
>>> +    decode_r_type_insn (SRLW, ival);
>>> +      else if (is_sra_insn(ival))
>>> +    decode_r_type_insn (SRA, ival);
>>> +      else if (is_sraw_insn(ival))
>>> +    decode_r_type_insn (SRAW, ival);
>>> +      else if (is_or_insn(ival))
>>> +    decode_r_type_insn (OR, ival);
>>> +      else if (is_and_insn(ival))
>>> +    decode_r_type_insn (AND, ival);
>>>         else if (is_lr_w_insn (ival))
>>> -    decode_r_type_insn (LR, ival);
>>> +    decode_r_type_insn (LR_W, ival);
>>>         else if (is_lr_d_insn (ival))
>>> -    decode_r_type_insn (LR, ival);
>>> +    decode_r_type_insn (LR_D, ival);
>>>         else if (is_sc_w_insn (ival))
>>> -    decode_r_type_insn (SC, ival);
>>> +    decode_r_type_insn (SC_W, ival);
>>>         else if (is_sc_d_insn (ival))
>>> -    decode_r_type_insn (SC, ival);
>>> +    decode_r_type_insn (SC_D, ival);
>>>         else if (is_ecall_insn (ival))
>>>       decode_i_type_insn (ECALL, ival);
>>>         else if (is_ld_insn (ival))
>>> @@ -1944,6 +2025,24 @@ riscv_insn::decode (struct gdbarch *gdbarch, 
>>> CORE_ADDR pc)
>>>         m_rd = decode_register_index (ival, OP_SH_CRS1S);
>>>         m_imm.s = EXTRACT_CITYPE_LUI_IMM (ival);
>>>       }
>>> +      else if (is_c_srli_insn (ival))
>>> +    decode_cb_type_insn (SRLI, ival);
>>> +      else if (is_c_srai_insn (ival))
>>> +    decode_cb_type_insn (SRAI, ival);
>>> +      else if (is_c_andi_insn (ival))
>>> +    decode_cb_type_insn (ANDI, ival);
>>> +      else if (is_c_sub_insn (ival))
>>> +    decode_ca_type_insn (SUB, ival);
>>> +      else if (is_c_xor_insn (ival))
>>> +    decode_ca_type_insn (XOR, ival);
>>> +      else if (is_c_or_insn (ival))
>>> +    decode_ca_type_insn (OR, ival);
>>> +      else if (is_c_and_insn (ival))
>>> +    decode_ca_type_insn (AND, ival);
>>> +      else if (is_c_subw_insn (ival))
>>> +    decode_ca_type_insn (SUBW, ival);
>>> +      else if (is_c_addw_insn (ival))
>>> +    decode_ca_type_insn (ADDW, ival);
>>>         else if (is_c_li_insn (ival))
>>>       decode_ci_type_insn (LI, ival);
>>>         /* C_SD and C_FSW have the same opcode.  C_SD is RV64 and 
>>> RV128 only,
>>> @@ -4404,51 +4503,138 @@ riscv_next_pc (struct regcache *regcache, 
>>> CORE_ADDR pc)
>>>     return next_pc;
>>>   }
>>>
>>> +/* Return true if INSN is not a control transfer instruction and is 
>>> allowed to
>>> +   appear in the middle of the lr/sc sequence.  */
>>> +
>>> +static bool
>>> +riscv_insn_is_non_cti_and_allowed_in_atomic_sequence(const struct 
>>> riscv_insn &insn)
>>> +{
>>> +  switch (insn.opcode ())
>>> +    {
>>> +    case riscv_insn::LUI:
>>> +    case riscv_insn::AUIPC:
>>> +    case riscv_insn::ADDI:
>>> +    case riscv_insn::ADDIW:
>>> +    case riscv_insn::SLTI:
>>> +    case riscv_insn::SLTIU:
>>> +    case riscv_insn::XORI:
>>> +    case riscv_insn::ORI:
>>> +    case riscv_insn::ANDI:
>>> +    case riscv_insn::SLLI:
>>> +    case riscv_insn::SLLIW:
>>> +    case riscv_insn::SRLI:
>>> +    case riscv_insn::SRLIW:
>>> +    case riscv_insn::SRAI:
>>> +    case riscv_insn::ADD:
>>> +    case riscv_insn::ADDW:
>>> +    case riscv_insn::SRAIW:
>>> +    case riscv_insn::SUB:
>>> +    case riscv_insn::SUBW:
>>> +    case riscv_insn::SLL:
>>> +    case riscv_insn::SLLW:
>>> +    case riscv_insn::SLT:
>>> +    case riscv_insn::SLTU:
>>> +    case riscv_insn::XOR:
>>> +    case riscv_insn::SRL:
>>> +    case riscv_insn::SRLW:
>>> +    case riscv_insn::SRA:
>>> +    case riscv_insn::SRAW:
>>> +    case riscv_insn::OR:
>>> +    case riscv_insn::AND:
>>> +      return true;
>>> +    }
>>> +
>>> +    return false;
>>> +}
>>> +
>>> +/* Return true if INSN is a direct branch insctruction.  */
>>> +
>>> +static bool
>>> +riscv_insn_is_direct_branch(const struct riscv_insn &insn)
>>> +{
>>> +  switch (insn.opcode ())
>>> +    {
>>> +    case riscv_insn::BEQ:
>>> +    case riscv_insn::BNE:
>>> +    case riscv_insn::BLT:
>>> +    case riscv_insn::BGE:
>>> +    case riscv_insn::BLTU:
>>> +    case riscv_insn::BGEU:
>>> +    case riscv_insn::JAL:
>>> +      return true;
>>> +    }
>>> +
>>> +    return false;
>>> +}
>>> +
>>>   /* We can't put a breakpoint in the middle of a lr/sc atomic 
>>> sequence, so look
>>>      for the end of the sequence and put the breakpoint there.  */
>>>
>>> -static bool
>>> -riscv_next_pc_atomic_sequence (struct regcache *regcache, CORE_ADDR 
>>> pc,
>>> -                   CORE_ADDR *next_pc)
>>> +static std::vector<CORE_ADDR>
>>> +riscv_deal_with_atomic_sequence (struct regcache *regcache, 
>>> CORE_ADDR pc)
>>>   {
>>>     struct gdbarch *gdbarch = regcache->arch ();
>>>     struct riscv_insn insn;
>>> -  CORE_ADDR cur_step_pc = pc;
>>> -  CORE_ADDR last_addr = 0;
>>> +  CORE_ADDR cur_step_pc = pc, next_pc;
>>> +  std::vector<CORE_ADDR> next_pcs;
>>> +  const int atomic_sequence_length = 16;
>>> +  bool found_valid_atomic_sequence = false;
>>> +  enum riscv_insn::opcode lr_opcode;
>>>
>>>     /* First instruction has to be a load reserved.  */
>>>     insn.decode (gdbarch, cur_step_pc);
>>> -  if (insn.opcode () != riscv_insn::LR)
>>> -    return false;
>>> -  cur_step_pc = cur_step_pc + insn.length ();
>>> +  lr_opcode = insn.opcode ();
>>> +  if (lr_opcode != riscv_insn::LR_D && lr_opcode != riscv_insn::LR_W)
>>> +    return {};
>>>
>>> -  /* Next instruction should be branch to exit.  */
>>> -  insn.decode (gdbarch, cur_step_pc);
>>> -  if (insn.opcode () != riscv_insn::BNE)
>>> -    return false;
>>> -  last_addr = cur_step_pc + insn.imm_signed ();
>>> -  cur_step_pc = cur_step_pc + insn.length ();
>>> +  /* A lr/sc sequence comprise at most 16 instructions placed 
>>> sequentially in memory. */
>>> +  for (int insn_count = 0; insn_count < atomic_sequence_length; 
>>> ++insn_count)
>>> +    {
>>> +      cur_step_pc += insn.length ();
>>> +      insn.decode (gdbarch, cur_step_pc);
>>>
>>> -  /* Next instruction should be store conditional.  */
>>> -  insn.decode (gdbarch, cur_step_pc);
>>> -  if (insn.opcode () != riscv_insn::SC)
>>> -    return false;
>>> -  cur_step_pc = cur_step_pc + insn.length ();
>>> +      /* The dynamic code executed between lr/sc can only contain 
>>> instructions
>>> +     from the base I instruction set, excluding loads, stores, 
>>> backward jumps,
>>> +     taken backward branches, JALR, FENCE, FENCE.I, and SYSTEM 
>>> instructions.
>>> +     If the C extension is supported, then compressed forms of the 
>>> aforementioned
>>> +     I instructions are also permitted.  */
>
> That ends up with some odd interactions with the Zc extensions.  I 
> opened https://github.com/riscv/riscv-isa-manual/issues/1157 to check.
>
Okay, let's track this concern there.
>>> +
>>> +      if (riscv_insn_is_non_cti_and_allowed_in_atomic_sequence (insn))
>>> +    {
>>> +      continue;
>>> +    }
>>> +      /* Look for a conditional branch instruction, check if it's 
>>> taken forward or not. */
>>> +      else if (riscv_insn_is_direct_branch (insn))
>>> +    {
>>> +      if (insn.imm_signed () > 0)
>>> +        next_pc = cur_step_pc + insn.imm_signed ();
>>
>> Oh I forgot to do a push_back here and my simple tests didn't catch
>> this. I'll send a V2 patch soon. Sorry.
>
V2 posted.
> I also think the branch detection logic here isn't quite right: the 
> LR/SC sequence only becomes unconstrained if the backwards branch is 
> taken at runtime.  I can't find any examples of people doing that, but 
> one could imagine some sort of spin-type patterns with an early retry 
> that do this sort of thing.
>
> The same logic applies to disallowed instructions, but I don't think 
> it's super likely those end up conditionally executed in LR/SC 
> sequences so maybe we can ignore that?
>
Yeah, it seems difficult to get the lr/sc sequence completely correct, 
so can we implement a version that is correct in most cases (as this 
patch does)?
>> Best regards,
>>
>> Yang
>>
>>> +      else
>>> +        break;
>>> +    }
>>> +      /* Look for a paired SC instruction which closes the atomic 
>>> sequence. */
>>> +      else if ((insn.opcode () == riscv_insn::SC_D && lr_opcode == 
>>> riscv_insn::LR_D)
>>> +           || (insn.opcode () == riscv_insn::SC_W && lr_opcode == 
>>> riscv_insn::LR_W))
>>> +    {
>>> +      found_valid_atomic_sequence = true;
>>> +    }
>>> +      else
>>> +    break;
>>> +    }
>>> +
>>> +  if (!found_valid_atomic_sequence)
>>> +    return {};
>>>
>>>     /* Next instruction should be branch to start.  */
>>>     insn.decode (gdbarch, cur_step_pc);
>>>     if (insn.opcode () != riscv_insn::BNE)
>>> -    return false;
>>> +    return {};
>>>     if (pc != (cur_step_pc + insn.imm_signed ()))
>>> -    return false;
>>> -  cur_step_pc = cur_step_pc + insn.length ();
>>> -
>>> -  /* We should now be at the end of the sequence.  */
>>> -  if (cur_step_pc != last_addr)
>>> -    return false;
>>> +    return {};
>>> +  cur_step_pc += insn.length ();
>>>
>>> -  *next_pc = cur_step_pc;
>>> -  return true;
>>> +  next_pc = cur_step_pc;
>>> +  next_pcs.push_back (next_pc);
>>> +  return next_pcs;
>>>   }
>>>
>>>   /* This is called just before we want to resume the inferior, if 
>>> we want to
>>> @@ -4458,14 +4644,14 @@ riscv_next_pc_atomic_sequence (struct 
>>> regcache *regcache, CORE_ADDR pc,
>>>   std::vector<CORE_ADDR>
>>>   riscv_software_single_step (struct regcache *regcache)
>>>   {
>>> -  CORE_ADDR pc, next_pc;
>>> -
>>> -  pc = regcache_read_pc (regcache);
>>> +  CORE_ADDR cur_pc = regcache_read_pc (regcache), next_pc;
>>> +  std::vector<CORE_ADDR> next_pcs
>>> +    = riscv_deal_with_atomic_sequence (regcache, cur_pc);
>>>
>>> -  if (riscv_next_pc_atomic_sequence (regcache, pc, &next_pc))
>>> -    return {next_pc};
>>> +  if (!next_pcs.empty ())
>>> +    return next_pcs;
>>>
>>> -  next_pc = riscv_next_pc (regcache, pc);
>>> +  next_pc = riscv_next_pc (regcache, cur_pc);
>>>
>>>     return {next_pc};
>>>   }


      reply	other threads:[~2023-11-02 17:55 UTC|newest]

Thread overview: 4+ messages / expand[flat|nested]  mbox.gz  Atom feed  top
2023-11-02 14:03 Yang Liu
2023-11-02 15:21 ` Yang Liu
2023-11-02 15:58   ` Palmer Dabbelt
2023-11-02 17:54     ` Yang Liu [this message]

Reply instructions:

You may reply publicly to this message via plain-text email
using any one of the following methods:

* Save the following mbox file, import it into your mail client,
  and reply-to-all from there: mbox

  Avoid top-posting and favor interleaved quoting:
  https://en.wikipedia.org/wiki/Posting_style#Interleaved_style

* Reply using the --to, --cc, and --in-reply-to
  switches of git-send-email(1):

  git send-email \
    --in-reply-to=bcd45a50-ec51-49fe-ae93-c19226f28d4a@iscas.ac.cn \
    --to=liuyang22@iscas.ac.cn \
    --cc=aburgess@redhat.com \
    --cc=gdb-patches@sourceware.org \
    --cc=palmer@dabbelt.com \
    --cc=simon.marchi@polymtl.ca \
    /path/to/YOUR_REPLY

  https://kernel.org/pub/software/scm/git/docs/git-send-email.html

* If your mail client supports setting the In-Reply-To header
  via mailto: links, try the mailto: link
Be sure your reply has a Subject: header at the top and a blank line before the message body.
This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox